Learning Negotiation Policies Using Ensemble-Based Drift Detection Techniques

In this work we compare drift detection techniques and we show how they can improve the performance of trade agents in multi-issue bilateral dynamic negotiations. In a dynamic negotiation the utility values and functions of trade agents can change on the fly. Intelligent trade agents must identify and take such drift in the competitors into account changing also the offer policies to improve the global utility throughout the negotiation. However, traditional learning mechanisms disregard possible changes in a competitor’s offer/counter-offer policy. In that case, the agent performance may decrease drastically. In our approach, a trade agent has a staff of weighted learner agents used to predict interesting offers. The staff uses the Dynamic Weighted Majority (DWM) algorithm to adapt itself creating, deleting and adapting staff members. The results obtained with the IB3 (Instance-based) learners and IB3-DWM learners show that ensemble methods like DWM are suitable for correctly identifying changes in agent negotiations.